This is the mail archive of the cygwin mailing list for the Cygwin project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Perl bug (was Re: [1.7] cygwin allows writing to readonly files)

2009/8/11 Reini Urban:
> 2009/8/10 Corinna Vinschen:
>> On Aug 10 20:11, Alexey Borzenkov wrote:
>>> On Mon, Aug 10, 2009 at 5:25 PM, Corinna
>>> Vinschen<> wrote:
>>> > That's a bug in your testsuite. ?I assume you're running the tests as
>>> > administrator, right? ?Administrators have the right to write to all
>>> > files, even R/O files, according to POSIX rules. ?Your test would fail
>>> > on Linux as well, if you're running it as root.
>>> Well, it's not my testsuite, but yes, I'm running under administrator
>>> account. But it makes me wonder, how does it work? Do you change ACLs
>>> temporarily?
>> No. ?It's a "user privileges" thingy. ?See
>> The SE_BACKUP_NAME and SE_RESTORE_NAME privileges are in the
>> administrator's user token, but they are not enabled by default. ?Cygwin
>> just enables them at startup time, if they are available in the user
>> token. ?Therefore, a Cygwin process has the usual POSIX-like permissions
>> for admin users. ?It's no magic which isn't available to any other
>> native Win32 application.
>>> Anyway, it means there is a bug in perl, because on Linux:
>>> root@kitsu:~# touch test.txt
>>> root@kitsu:~# chmod 0444 test.txt
>>> root@kitsu:~# perl -e 'print "writable\n" if -w "test.txt"'
>>> writable
>>> On Cygwin 1.7 perl thinks that the file is not writable.
>> Indeed. ?Checking with strace I found that the test is the same on Linux
>> and Cygwin. ?In both cases perl uses stat(), and the returned permissions
>> are the same (0444). ?Further experimenting shows that perl has a
>> hardcoded uid == 0 test which must obviously fail on Cygwin. ?If I change
>> the user's uid to 0, the string "writable" is printed by the above command.
>> That's a bug in perl. ?There are other OSes out there which have
>> root-like permissions for non-0 uids. ?Perl should use the access()
>> function to check for read/write/execute permissions, which always
>> returns the correct result independent of the uid of the current user.
> Thanks.
> I'll carry it along to p5p, but it will definitely not appear in
> upstream 5.10.1
> because this gate is already closed.
> Even a horrible performance problem with
> File::Spec::Cygwin::case_tolerant was not fixed.
> But I work on a fix to be included in blead and in my cygwin package.

Bug confirmed too early. It's actually defined and described this way.
access() is not used for performance reasons. See perldoc perlfunc -X

If you are using ACLs, there is a pragma called C<filetest> that may
produce more accurate results than the bare stat() mode bits.
When under the C<use filetest 'access'> the above-mentioned filetests
will test whether the permission can (not) be granted using the
access() family of system calls.  Also note that the C<-x> and C<-X> may
under this pragma return true even if there are no execute permission
bits set (nor any extra execute permission ACLs).  This strangeness is
due to the underlying system calls' definitions. Note also that, due to
the implementation of C<use filetest 'access'>, the C<_> special
filehandle won't cache the results of the file tests when this pragma is
in effect.  Read the documentation for the C<filetest> pragma for more

$ ./perl -e 'print "writable\n" if -w "test.txt"'

$ ./perl -e 'use filetest "access"; print "writable\n" if -w "test.txt"'

I can turn on access checks easily for CYGWIN but cygwin perl is already
slow enough, so I will not do that.

Changing the uid==0 check to check the Administrators gid is more promising.
--- doio.c.orig 2009-04-18 19:17:04.000000000 +0200
+++ doio.c      2009-08-11 04:46:09.343750000 +0200
@@ -1918,7 +1918,11 @@
      return (mode & statbufp->st_mode) ? TRUE : FALSE;

 #else /* ! DOSISH */
+# ifndef __CYGWIN__
     if ((effective ? PL_euid : PL_uid) == 0) { /* root is special */
+# else
+    if ((effective ? PL_egid : PL_gid) == 544) { /* member of
Administrators? */
+# endif
        if (mode == S_IXUSR) {
            if (statbufp->st_mode & 0111 || S_ISDIR(statbufp->st_mode))
                return TRUE;

but this didn't help me, because Administrators is not my first group.
So I call this a known limitation for all ACL aware filesystems.
Reini Urban 

Problem reports:
Unsubscribe info:

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]